Update core/extract.py
Browse files- core/extract.py +139 -26
core/extract.py
CHANGED
|
@@ -1,9 +1,60 @@
|
|
|
|
|
| 1 |
from __future__ import annotations
|
| 2 |
-
import json
|
| 3 |
-
from typing import
|
| 4 |
-
from .openai_client import get_client, VISION_MODEL, TEXT_MODEL
|
| 5 |
|
| 6 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 7 |
与えられた決算書(画像またはテキスト)から、次の厳密な JSON 構造のみを日本語の単位なし・半角数値で返してください。分からない項目は null。
|
| 8 |
{
|
| 9 |
"company": {"name": null},
|
|
@@ -24,42 +75,104 @@ SYSTEM_JSON = """あなたは有能な財務アナリストです。
|
|
| 24 |
}
|
| 25 |
"""
|
| 26 |
|
| 27 |
-
def
|
| 28 |
-
client =
|
| 29 |
if images:
|
| 30 |
-
content = [{"type": "text", "text":
|
| 31 |
if company_hint:
|
| 32 |
content.append({"type": "text", "text": f"会社名の候補: {company_hint}"})
|
| 33 |
for im in images:
|
| 34 |
-
content.append({"type": "input_image", "image_url": f"data:image/png;base64,{
|
| 35 |
-
# 上のデータ URI 生成は UI 側で行うためここでは未使用
|
| 36 |
-
# (UIでdata:image/png;base64,xxxを組む実装に合わせる場合は差し替え)
|
| 37 |
-
pass
|
| 38 |
-
# 実運用では UI 側で Vision を呼ぶ形にせず、ここで共通化
|
| 39 |
-
if images:
|
| 40 |
-
content = [{"type":"text","text":SYSTEM_JSON}]
|
| 41 |
-
for im in images:
|
| 42 |
-
import base64
|
| 43 |
-
content.append({"type":"input_image","image_url":f"data:image/png;base64,{base64.b64encode(im).decode('utf-8')}"})
|
| 44 |
resp = client.chat.completions.create(
|
| 45 |
-
model=
|
| 46 |
messages=[
|
| 47 |
-
{"role":"system","content":"返答は必ず有効な JSON
|
| 48 |
-
{"role":"user","content":content},
|
| 49 |
],
|
| 50 |
-
response_format={"type":"json_object"},
|
| 51 |
temperature=0.1,
|
| 52 |
)
|
| 53 |
return json.loads(resp.choices[0].message.content)
|
| 54 |
else:
|
| 55 |
-
prompt = f"{
|
| 56 |
resp = client.chat.completions.create(
|
| 57 |
-
model=
|
| 58 |
messages=[
|
| 59 |
-
{"role":"system","content":"返答は必ず有効な JSON
|
| 60 |
-
{"role":"user","content":prompt},
|
| 61 |
],
|
| 62 |
-
response_format={"type":"json_object"},
|
| 63 |
temperature=0.1,
|
| 64 |
)
|
| 65 |
return json.loads(resp.choices[0].message.content)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# core/extract.py
|
| 2 |
from __future__ import annotations
|
| 3 |
+
import os, io, base64, json, shutil
|
| 4 |
+
from typing import List, Tuple, Dict, Any
|
|
|
|
| 5 |
|
| 6 |
+
import pandas as pd
|
| 7 |
+
import pdfplumber
|
| 8 |
+
from pdf2image import convert_from_path
|
| 9 |
+
from openai import OpenAI
|
| 10 |
+
|
| 11 |
+
# ==== モデル指定(環境変数で変更可) ====
|
| 12 |
+
OPENAI_MODEL_VISION = os.environ.get("OPENAI_VISION_MODEL", "gpt-4o-mini")
|
| 13 |
+
OPENAI_MODEL_TEXT = os.environ.get("OPENAI_TEXT_MODEL", "gpt-4o-mini")
|
| 14 |
+
|
| 15 |
+
# ==== 共通ユーティリティ ====
|
| 16 |
+
def _b64(b: bytes) -> str:
|
| 17 |
+
return base64.b64encode(b).decode("utf-8")
|
| 18 |
+
|
| 19 |
+
def _client() -> OpenAI:
|
| 20 |
+
"""
|
| 21 |
+
OpenAI公式 SDK v1 系。httpx との互換のため requirements は httpx==0.27.* を推奨。
|
| 22 |
+
"""
|
| 23 |
+
key = os.environ.get("OPENAI_API_KEY")
|
| 24 |
+
if not key:
|
| 25 |
+
raise RuntimeError("OPENAI_API_KEY が未設定です(Spaces → Settings → Variables and secrets)。")
|
| 26 |
+
# proxies は渡さない(互換性エラーを避ける)
|
| 27 |
+
return OpenAI(api_key=key, timeout=30)
|
| 28 |
+
|
| 29 |
+
# ==== PDF 読み込み ====
|
| 30 |
+
def _pdf_to_images(path: str, dpi: int = 220, max_pages: int = 6) -> List[bytes]:
|
| 31 |
+
"""
|
| 32 |
+
Poppler系バイナリ(pdftoppm/pdftocairo)が必要です。Spaces なら packages.txt に
|
| 33 |
+
`poppler-utils` を入れておくと安定します。
|
| 34 |
+
"""
|
| 35 |
+
imgs: List[bytes] = []
|
| 36 |
+
pages = convert_from_path(path, dpi=dpi, fmt="png")
|
| 37 |
+
for i, p in enumerate(pages):
|
| 38 |
+
if i >= max_pages:
|
| 39 |
+
break
|
| 40 |
+
buf = io.BytesIO()
|
| 41 |
+
p.save(buf, format="PNG")
|
| 42 |
+
imgs.append(buf.getvalue())
|
| 43 |
+
return imgs
|
| 44 |
+
|
| 45 |
+
def _pdf_to_text(path: str, max_chars: int = 15000) -> str:
|
| 46 |
+
out: List[str] = []
|
| 47 |
+
with pdfplumber.open(path) as pdf:
|
| 48 |
+
for i, page in enumerate(pdf.pages):
|
| 49 |
+
t = (page.extract_text() or "").strip()
|
| 50 |
+
if t:
|
| 51 |
+
out.append(f"[page {i+1}]\n{t}")
|
| 52 |
+
if sum(len(x) for x in out) > max_chars:
|
| 53 |
+
break
|
| 54 |
+
return "\n\n".join(out)[:max_chars]
|
| 55 |
+
|
| 56 |
+
# ==== LLM へ渡す JSON 指示 ====
|
| 57 |
+
_SYSTEM_JSON = """あなたは有能な財務アナリストです。
|
| 58 |
与えられた決算書(画像またはテキスト)から、次の厳密な JSON 構造のみを日本語の単位なし・半角数値で返してください。分からない項目は null。
|
| 59 |
{
|
| 60 |
"company": {"name": null},
|
|
|
|
| 75 |
}
|
| 76 |
"""
|
| 77 |
|
| 78 |
+
def _extract_with_llm(images: List[bytes] | None, text_blob: str | None, company_hint: str) -> Dict[str, Any]:
|
| 79 |
+
client = _client()
|
| 80 |
if images:
|
| 81 |
+
content = [{"type": "text", "text": _SYSTEM_JSON}]
|
| 82 |
if company_hint:
|
| 83 |
content.append({"type": "text", "text": f"会社名の候補: {company_hint}"})
|
| 84 |
for im in images:
|
| 85 |
+
content.append({"type": "input_image", "image_url": f"data:image/png;base64,{_b64(im)}"})
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 86 |
resp = client.chat.completions.create(
|
| 87 |
+
model=OPENAI_MODEL_VISION,
|
| 88 |
messages=[
|
| 89 |
+
{"role": "system", "content": "返答は必ず有効な JSON オブジェクトのみ。説明を含めない。"},
|
| 90 |
+
{"role": "user", "content": content},
|
| 91 |
],
|
| 92 |
+
response_format={"type": "json_object"},
|
| 93 |
temperature=0.1,
|
| 94 |
)
|
| 95 |
return json.loads(resp.choices[0].message.content)
|
| 96 |
else:
|
| 97 |
+
prompt = f"{_SYSTEM_JSON}\n\n以下は決算書のテキストです。上記の JSON だけを返してください。\n\n{text_blob or ''}"
|
| 98 |
resp = client.chat.completions.create(
|
| 99 |
+
model=OPENAI_MODEL_TEXT,
|
| 100 |
messages=[
|
| 101 |
+
{"role": "system", "content": "返答は必ず有効な JSON オブジェクトのみ。"},
|
| 102 |
+
{"role": "user", "content": prompt},
|
| 103 |
],
|
| 104 |
+
response_format={"type": "json_object"},
|
| 105 |
temperature=0.1,
|
| 106 |
)
|
| 107 |
return json.loads(resp.choices[0].message.content)
|
| 108 |
+
|
| 109 |
+
# ==== JSON <-> DataFrame ====
|
| 110 |
+
def _fin_to_df(fin: Dict[str, Any]) -> pd.DataFrame:
|
| 111 |
+
rows: List[Dict[str, Any]] = []
|
| 112 |
+
def add(cat: str, d: Dict[str, Any] | None):
|
| 113 |
+
for k, v in (d or {}).items():
|
| 114 |
+
rows.append({"category": cat, "item": k, "value": v})
|
| 115 |
+
add("balance_sheet", fin.get("balance_sheet"))
|
| 116 |
+
add("income_statement", fin.get("income_statement"))
|
| 117 |
+
add("cash_flows", fin.get("cash_flows"))
|
| 118 |
+
return pd.DataFrame(rows, columns=["category", "item", "value"])
|
| 119 |
+
|
| 120 |
+
# ==== 公開 API:parse_pdf ====
|
| 121 |
+
def parse_pdf(files: List[str], company: str = "", force_ocr: bool = False) -> Tuple[Dict[str, Any], pd.DataFrame]:
|
| 122 |
+
"""
|
| 123 |
+
入力: PDFファイルパスの配列
|
| 124 |
+
出力: (抽出JSON辞書, 表編集用DataFrame)
|
| 125 |
+
方針:
|
| 126 |
+
- まず PDF→画像化して Vision で抽出(poppler が無い/失敗なら例外)
|
| 127 |
+
- 画像抽出が失敗したらテキスト抽出→Textモデルで抽出
|
| 128 |
+
- `force_ocr=True` の場合は常に画像→Vision を試みる
|
| 129 |
+
"""
|
| 130 |
+
if not files:
|
| 131 |
+
raise ValueError("PDF が指定されていません。")
|
| 132 |
+
|
| 133 |
+
# 1) 画像化(複数PDFを順に)
|
| 134 |
+
images: List[bytes] = []
|
| 135 |
+
if force_ocr:
|
| 136 |
+
for p in files:
|
| 137 |
+
images += _pdf_to_images(p, dpi=220, max_pages=6)
|
| 138 |
+
else:
|
| 139 |
+
# 画像化を試して、ダメならテキストにフォールバック
|
| 140 |
+
try:
|
| 141 |
+
for p in files:
|
| 142 |
+
images += _pdf_to_images(p, dpi=220, max_pages=6)
|
| 143 |
+
except Exception:
|
| 144 |
+
images = []
|
| 145 |
+
|
| 146 |
+
# 2) Vision / Text のいずれかで抽出
|
| 147 |
+
try:
|
| 148 |
+
if images:
|
| 149 |
+
fin = _extract_with_llm(images, None, company or "")
|
| 150 |
+
else:
|
| 151 |
+
# テキスト抽出
|
| 152 |
+
text_blob = ""
|
| 153 |
+
for p in files:
|
| 154 |
+
text_blob += _pdf_to_text(p) + "\n\n"
|
| 155 |
+
fin = _extract_with_llm(None, text_blob, company or "")
|
| 156 |
+
except Exception as e:
|
| 157 |
+
# LLM失敗時も最後にテキスト抽出で最低限の骨格を返す
|
| 158 |
+
text_blob = ""
|
| 159 |
+
for p in files:
|
| 160 |
+
try:
|
| 161 |
+
text_blob += _pdf_to_text(p) + "\n\n"
|
| 162 |
+
except Exception:
|
| 163 |
+
pass
|
| 164 |
+
fin = {
|
| 165 |
+
"company": {"name": company or None},
|
| 166 |
+
"period": {"start_date": None, "end_date": None},
|
| 167 |
+
"balance_sheet": {"total_assets": None, "total_liabilities": None, "total_equity": None,
|
| 168 |
+
"current_assets": None, "fixed_assets": None,
|
| 169 |
+
"current_liabilities": None, "long_term_liabilities": None},
|
| 170 |
+
"income_statement": {"sales": None, "cost_of_sales": None, "gross_profit": None,
|
| 171 |
+
"operating_expenses": None, "operating_income": None,
|
| 172 |
+
"ordinary_income": None, "net_income": None},
|
| 173 |
+
"cash_flows": {"operating_cash_flow": None, "investing_cash_flow": None, "financing_cash_flow": None},
|
| 174 |
+
"_fallback_note": f"LLM抽出に失敗したため簡易骨格のみ返却(理由: {type(e).__name__})"
|
| 175 |
+
}
|
| 176 |
+
|
| 177 |
+
df = _fin_to_df(fin)
|
| 178 |
+
return fin, df
|